llama : refactor GGUF constants into static maps
4ed3469c
llama : check if model architecture is known
8bd7f06b
llama : refactor llama_model_load_internal()
3057d6a6
gguf : add KV constant maps
3c025a6d
Merge branch 'master' into falcon
b19c6e46
llm : read arch-specific KVs
9f28f737
convert : add dummy scores + types
d1b3b95d
falcon : load tensor data (CPU only)
2f3c80a8
llama : fix loading progress bar
5c5413dc
llama : add arch member to llama_model
085228e1
falcon : CPU inference working
3c7c325b
falcon : support non-40B models
2d58444d
falcon : minor
0ec27ad6
llama : minor updates
7bbbf38c
convert-falcon-hf-to-gguf.py : fix special token mapping
9853f2cf
llama.cpp : llama default UNK token = id 0
ffa5099c
llama.cpp : fix bpe tokenizer
a95ae752
llama.cpp : fix the fix of bpe tokenizer
d561b7f7
ggml : pass eps to ggml_norm
e3c52bd9
metal : implement RoPE (mode = 2) + avoid ggml_repeat
99bb2607
ggml : ggml_repeat always creates new tensor
af4bbcc8
falcon : copy-paste self-attention from LLaMA
b34ab740
metal : print extra compute pipeline info
a0dc47a5
falcon : minor changes (still chasing the Metal problem)
e2d23bed
llama.cpp : fix linefeed token
b693000c
metal : fix GELU kernel numerical stability by using precise::tanh
0a85ae73
metal : temporary workaround for the concurrency optimization bug
854ae5d0
falcon : add CUDA offloading (#2739)
e7299656
llama : better model naming and size reporting
176ea716
Merge branch 'master' into falcon
6938c5f4
llama : prep new tokenizer support
c3f8a6e4
llama : advanced BPE tokenizer based on ggllm.cpp imlpementation
3bfb7206
llama : remove oboslete comment
2424e1d0
common : remove obsolete BPE API + disable test-tokenizer-1
596e1094
llama : revert BPE special-case in llama_byte_to_token()
f8ee54bd
cuda : add TODOs for RoPE NeoX implementation
8c6d3939
llama : default special tokens based on vocab type
630d8b40
perplexity : add log for start of tokenization
fae8faa1
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub